Neural Networks to Infer Traditional Chinese Medicine Prescriptions from Indications

Ith increasing digitization of Chinese medicine-related books and extraction and analysis of the ingredients in herbs, it now becomes feasible to use big data analysis and deep learning techniques to learn the regularities from previous vague experience and knowledge of traditional Chinese medicine. We combine the Compendium of Materia Medica, Traditional Chinese Medicine Integrated Database (TCMID) and Traditional Chinese Medicine Systems Pharmacology Database (TCMSPD) and uses a pre-trained ensemble convolutional neural networks to infer Chinese medicine prescriptions from Chinese medicine indications. We constructed multiple biological networks including indications, target proteins and chemical compounds, and inferred the ingredients using a random walk algorithm from indications, and the potential Chinese medicine prescriptions are generated by a combination of herbs that cover the inferred ingredients. A pre-trained ensemble CNN is used to filter out unlikely prescriptions. Even under extreme incomplete information of the domain knowledge, the blind evaluation by human experts on the prescriptions proposed by our system being categorized as “suitable” or “very suitable” against “not suitable” and “very unsuitable” is overall 38.00%.